All posts
8 min read

OpenClaw vs Choreo: own it all, or skip the ops

If you're researching always-on AI assistants in 2026, you've met OpenClaw — the open-source project that went from launch to over 250,000 GitHub stars in about two months. We build Choreo, a platform in the same category, so read this knowing where it comes from. But that is also why it can be specific: these are the two honest architectures for an agent that works while you sleep — run it yourself and own everything, or run it on a platform that was built multi-tenant and isolated from the first commit — and which one is right for you is a real question with a real answer.

What OpenClaw is, in its own words

OpenClaw (born Clawdbot in late 2025, renamed in January 2026) is a free, open-source AI agent framework you run on your own Mac, Windows or Linux machine. Per its own site: you can talk to it on WhatsApp, Telegram, Discord, Slack, Signal, iMessage "or any of its 29 channels"; it browses the web, fills forms, reads and writes files, and runs shell commands — "full access or sandboxed, your choice"; and "state lives on your machine, not a vendor cloud." A heartbeat daemon makes it proactive rather than chat-shaped, and a community skill registry (ClawHub) lets it extend itself. It deserves its stars: it proved, loudly, that people want an agent that lives in their messaging apps and keeps working — the same conviction this category is built on.

Where OpenClaw wins

As of August 2026, per its own materials — and these are real wins:

  • It's free and it's yours. Open source, no subscription, auditable code, and your data stays on hardware you control. If sovereignty is the requirement, nothing hosted competes with that.
  • 29 channels to our 10. Including iMessage and Signal, which Choreo does not have. If your life runs on iMessage, that alone may decide it.
  • Ecosystem velocity. A community that size ships skills, integrations and fixes at a pace no vendor roadmap matches.
  • Total configurability. "Full access" mode exists. When you genuinely want an agent with the keys to your whole machine, OpenClaw will hand them over. We won't.

The security question, with receipts

Here is the part most comparisons soften. OpenClaw's power comes from running with real access on a real machine, and its security model depends on every individual operator configuring it well. At the scale of a quarter-million-star project, that assumption met reality — all of the following is third-party reporting, not our claim:

  • Researchers at Wiz found over 135,000 OpenClaw instances reachable on the public internet in February 2026; SecurityScorecard separately reported 40,000+ exposed instances with roughly a third flagged vulnerable.
  • Published CVEs have covered command injection, SSRF, path traversal and prompt-injection-driven code execution, and security researchers tracking ClawHub counted around 1,400 malicious skills — credential stealers and reverse shells wearing productivity names.
  • Microsoft's security team wrote in February 2026 that "it is not appropriate to run it on a standard personal or corporate machine," and China's MIIT issued a formal advisory about default configurations the same month.

None of this makes OpenClaw bad software — it makes it software whose blast radius is your machine, your files and your accounts, governed by configuration you maintain. The hardening guides all say the same thing: sandbox it, isolate its network, never give it raw credentials, log everything it does. Choreo is what you get when that checklist is the architecture instead of the homework:

  • Every run is sandboxed with gVisor, a userspace kernel between agent code and the host.
  • Containers never hold your keys — a gateway makes credentialed calls on the agent's behalf, and outbound traffic is checked against allowlists and pinned to the address that was checked.
  • Every capability is allow/ask/deny per agent, and a sub-agent can only ever be narrower than its parent.
  • Every model, tool and connector call lands in an append-only audit trail with its cost.

And there is no instance for you to accidentally expose, because there is no instance you operate — the details are on the trust page, including the limits we admit to.

The capability comparison, briefly

Beyond hosting and security, the platforms diverge on shape. OpenClaw is one assistant, deeply yours, on one machine. Choreo is teams of agents on persistent projects: each agent picks its own runtime — including OpenAI's Codex and Anthropic's Claude Code run as-is, side by side — with its own model, budget and permissions. Scheduled runs follow strict unattended rules (anything set to "ask" is denied, and email sending, credentials, file deletion and computer control are never grantable unattended). The phone line answers live, in production. Memory has a published recall formula and an editable store. And billing is whichever of three you prefer: platform credits, your own API keys at no markup, or the Claude Code / ChatGPT subscription you already pay for.

Who should pick which

Pick OpenClaw if you enjoy operating software, want full ownership of data and code, need iMessage or Signal, and are genuinely prepared to do the hardening work — isolated hardware, locked-down network, vetted skills, and the patch cadence of a fast-moving project. Plenty of people fit that description, and for them it is a remarkable thing to run. Pick Choreo if you want the same always-on, message-it-anywhere agent life with the security model handled by architecture rather than by your weekend — or if the thing you actually want is a team of scoped agents rather than one assistant with your keys. The honest test costs nothing: run both side by side for a week — no card, no server — and let whichever one already did the work by Tuesday morning decide.

See it for yourself — free, $2 of credits once you confirm your email, no card.

Start free